Re: [PATCH 2/2] ARM: dts: exynos: use Exynos5420 dedicated USB2 PHY compatible

[Arnd Bergmann <arnd@kernel.org>]

On Wed, Jan 27, 2021 at 8:59 AM Krzysztof Kozlowski <krzk@kernel.org> wrote:
> On Tue, Jan 26, 2021 at 11:44:26PM +0100, Arnd Bergmann wrote:
> > On Fri, Nov 20, 2020 at 12:10 PM Marek Szyprowski <m.szyprowski@samsung.com> wrote:

> > > It won't work easily with both compatibles, because in the 5420 variant
> > > I've also changed the PHY indices (5420 has no device and second hsic
> > > phy). IMHO the dts change can wait for the next release.
> >
> > I see this made it into the pull request now, but I had not been aware
> > of the change earlier, and I'm slightly annoyed to have received it this
> > way:
> >
> > - This is clearly an incompatible change to the dtb, and you all
> >    noticed that because it would cause a bisection problem. As
> >    a general rule, if a dts change does not work across bisection,
> >    we should not merge it at all, because it causes problems for
> >    anyone with external dts or dtb files.
>
> Hi Arnd,
>
> No, it does not create a bisection problem. The driver change adding new
> compatible is already in v5.11-rc1.

What I meant is that you knew there would be a bisection problem
if you had not delayed this patch.

> > - It would likely have been possible to define the new binding in
> >   a backward-compatible way. I don't see a reason why the index
> >   values in the binding had to change here, other than a slight
> >   inconvenience for the driver.
>
> It does not matter since it's a new compatible and old one is not
> affected. Nothing got broken before this patch, nothing got broken after
> applying it via samsung-soc. No backwards compatibility is affected.
>
> > - If the change was really unavoidable, I would have expected
> >   a long explanation about why it had to be done in both the
> >   commit message and in the tag description for the pull
> >   request.
> >
> > I've dropped the pull request for now, maybe this can still
> > be sorted out with another driver change that makes the
> > new compatible string backward-compatible.
>
> It's a different hardware. New hardware does not have to be compatible
> with old hardware. However old DTB is still doing fine (although with
> the original issue not fixed).

There are around ten boards including this file, and most (maybe all)
of them are not newly added machines, so there is a good chance
that there are existing users. You are right that you took care that
the combination of an old dtb with a new kernel would not be any
worse than before, and that is good. What is however missing is
the consideration of the reverse: If anyone wants to dual-boot between
old and new kernels, they are stuck with the old dtb and is missing
the bugfix along with any additional changes that may get added
in the future.

The same is true if there are any non-Linux operating systems running
on these. For instance, FreeBSD runs on Peach Pit, and if they
were using the old dtb from Linux (I have not checked if they
were compatible before this change), then booting with the latest
dtb from Linux will require the same changes to their driver to avoid
a regression.

I can live with an explanation of "we've looked at all the alternatives
and decided to break old kernels with new dtbs in this particular
case because ...", but I don't like the idea of silently changing dts
in a way that breaks using them with anything but the latest kernel
and arguing that it's not even worth debating.

       Arnd